Synthetic or licensed training data: does AI still need real company records?
Short answer. Synthetic data is generated by a program or a model rather than collected from real activity. It is cheap to make and useful for filling gaps and protecting privacy, but a model trained mostly on generated output loses the rare cases real data contains. Labs use both, and real records from working companies remain the reference that synthetic sets are built from and checked against.
General information, not legal advice.
What synthetic data is
Synthetic data covers anything made rather than observed: text written by a model, simulated sensor readings, invented customer tables with realistic patterns. Some is generated from scratch by rules. Most modern synthetic text is produced by an existing model, often prompted with real examples so the output resembles them.
It has real uses. It can multiply a small set of rare cases, create practice tasks for an agent, or stand in for records that cannot leave a company. NIST lists generating synthetic data from identified records as one of the ways an agency can share data without exposing the people in it.
Where it falls short
A 2024 paper in Nature trained language models repeatedly on text produced by the previous generation. Over a few rounds the models lost the unusual parts of the original distribution and drifted toward repetitive, narrow output. The authors called this model collapse, and they concluded that training on generated data is possible only with careful filtering and a steady supply of human data.
There is a simpler limit too. A model can only generate what it already knows. It cannot invent how a particular freight dispatcher handles a missed pickup, or how an engineering firm argued over a change order, if no such records were in its training. Synthetic data reshapes knowledge; it does not add the missing kind.
Why real records keep their value
Real operating records carry the details that make work hard: incomplete information, wrong first guesses, exceptions, and the reason a decision went one way. Those details are what agents need to learn, and they are what a generated set tends to smooth away. Labs also use real records as the measuring stick, comparing a synthetic set against them to see whether it is realistic.
As more of the public web fills with model output, verified human records become scarcer in relative terms. That is one reason licensed data from companies, with a documented origin, has become a market at all.
How the two work together
In practice a lab may license a few years of real support tickets, generate many variations of the hardest cases, and test the result against held-back real tickets. The licensed set is the seed and the benchmark. For a company, this means synthetic data rarely replaces a license; it often depends on one.
What this means for a company with records
A company sometimes hears that synthetic data will make its archive worthless. The research points the other way. Generators need real examples to imitate, and labs need real examples to test whether generated sets are any good. Records that show rare situations, long histories and decisions with their reasons attached are the hardest thing for a generator to fake.
What synthetic data does change is the value of plain, repetitive records. If a generator can already produce convincing routine emails, a lab will pay little for more of them. The records worth licensing are the ones that capture judgement: escalations, disputes, change requests, failed attempts and how they were fixed. A useful exercise is to look through your own systems for those threads before talking to any buyer.
Follow-up questions
Could a lab generate my data instead of licensing it?
It can generate data that looks similar on the surface. It cannot reproduce the specific cases, decisions and history in your systems without having seen something like them, which is the gap a license fills.
Is synthetic data free of privacy risk?
Not automatically. A generator trained on personal records can leak details from them, so synthetic sets made from real data still need a privacy review.
Do labs say which of their data is synthetic?
Some disclose it in model documentation and some laws now ask for summaries of training data, but practice varies. Ask a buyer how your records will be combined with generated data.
Didn't find your question? Ask us directly.